Objective:

To develop a stacking ensemble model that integrates 3D intratumoral heterogeneity (3D ITH) scores with clinicoradiologic features for accurate preoperative prediction of visceral pleural invasion (VPI) in lung adenocarcinoma (LUAD), addressing the critical need for improved preoperative assessment.

Key Findings:

• The stacking ensemble classifier achieved the highest AUC (0.878) for preoperative prediction of VPI in LUAD, indicating superior model performance.
• 3D ITH score was identified as the most influential predictor, followed by nodule size and CT density.
• The stacking ensemble model outperformed the conventional radiomics signature (AUC = 0.841) and clinicoradiologic comparative model (AUC = 0.776), demonstrating its effectiveness.

Interpretation:

The integration of 3D ITH scores with clinicoradiologic features provides strong discrimination for predicting VPI in LUAD, suggesting its potential utility in preoperative risk stratification and treatment planning, which could significantly impact patient outcomes.

Limitations:

• Retrospective design may introduce selection bias, potentially affecting the generalizability of the findings.
• External validation limited to one center, which may not represent broader populations.
• Dependence on the quality of imaging data and radiomic feature extraction, which could influence model accuracy.

Conclusion:

The model integrating 3D ITH scores with clinicoradiologic features demonstrates strong predictive capability for VPI in LUAD, supporting its role in enhancing preoperative assessment and individualized treatment strategies.